autopas::AutoTuner Class Reference

This class manages all logic related to the auto tuning mechanic. More...

#include <AutoTuner.h>

Public Types
using	TuningStrategiesListType = std::vector< std::unique_ptr< TuningStrategyInterface > >
	Type for the member holding all tuning strategies.
using	SearchSpaceType = std::set< Configuration >
	Type for the search space holding all possible configurations.

Public Member Functions
	AutoTuner (TuningStrategiesListType &tuningStrategies, const SearchSpaceType &searchSpace, const AutoTunerInfo &autoTunerInfo, unsigned int rebuildFrequency, const std::string &outputSuffix)
	Constructor for the AutoTuner that generates all configurations from the given options.
AutoTuner &	operator= (AutoTuner &&other) noexcept
	Move assignment operator.
	~AutoTuner ()
	Default destructor defined in .cpp as it's too big to be inlined.
void	forceRetune ()
	Force the internal tuner to enter a new tuning phase upon the next call to computeInteractions().
const TuningMetricOption &	getTuningMetric () const
	Getter for the primary metric for tuning.
void	receiveLiveInfo (const LiveInfo &liveInfo, bool isStartOfTuningPhase)
	Pass live info on to all tuning strategies.
bool	needsLiveInfo () const
	Returns true if the AutoTuner needs live info.
bool	needsDomainSimilarityStatistics () const
	Returns true if the AutoTuner needs domain similarity statistics.
bool	willRebuildNeighborLists () const
	Returns whether rebuildNeighborLists() should be triggered in the next iteration.
const Configuration &	getCurrentConfig () const
	Get the currently selected configuration.
Configuration	rejectConfig (const Configuration &rejectedConfig, bool indefinitely, size_t tuningPhase)
	Tell the tuner that the given config is not applicable.
bool	searchSpaceIsTrivial () const
	Indicator function whether the search space consists of exactly one configuration.
bool	searchSpaceIsEmpty () const
	Indicator function whether the search space has no configurations in it.
void	logTuningResult (long tuningTime, size_t currentIteration) const
	After a tuning phase has finished, write the result to a file.
bool	initEnergy ()
	Initialize pmt sensor.
bool	resetEnergy ()
	Reset the rapl meter to prepare for a new measurement.
std::tuple< double, double, double, long >	sampleEnergy ()
	Take an energy measurement.
void	addMeasurement (long sampleRebuild, long sampleTraverseParticles, bool neighborListRebuilt, size_t iteration, size_t tuningPhase)
	Save the runtime of a given traversal.
void	addDomainSimilarityStatistics (double pdBinDensityStdDev, double pdBinMaxDensity)
	Adds domain similarity statistics to a vector of measurements, which can be smoothed for use in MPI Tuning to find similar domains.
const std::vector< Configuration > &	getConfigQueue () const
	Getter for the current queue of configurations.
const std::vector< std::unique_ptr< TuningStrategyInterface > > &	getTuningStrategies () const
	Get the list of tuning strategies that are used.
bool	inTuningPhase () const
	Indicate if the tuner considers itself currently in a tuning phase according to its internal counters.
bool	inLastTuningIteration () const
	Indicates if the tuner is in the last iteration of the tuning phase.
const EvidenceCollection &	getEvidenceCollection () const
	Getter for the internal evidence collection.
bool	canMeasureEnergy () const
	Returns whether the AutoTuner can take energy measurements.
void	setRebuildFrequency (double rebuildFrequency)
	Sets the _rebuildFrequency.
void	checkEarlyStoppingCondition (size_t tuningPhase)
	Checks whether the current configuration performs so poorly that it shouldn't be resampled further within this tuning phase.
bool	tuneConfiguration (size_t currentIteration, size_t tuningPhase, bool isStartOfTuningPhase)
	Tune available algorithm configurations.
void	setOptimalConfiguration (const Configuration &optimalConfig)
	Pushes the provided config into the _configQueue and ends tuning immediately.
const std::set< Configuration > &	getSearchSpace () const
	Return the search space as a const reference.

Detailed Description

This class manages all logic related to the auto tuning mechanic.

This involves:

• Managing the search space and configQueue.
• Managing and triggering all active tuning strategies.
• Managing all collected evidence and passing it to the strategies.

The tuner can be in one of two states. If it currently should look for a new optimum, it is in the so-called tuning phase. During a tuning phase, for each Configuration, multiple measurements can be taken, which are called samples. To reduce noise, the samples for one configuration are then condensed to one value for the current tuning phase, called evidence. The evidences are handed on to a tuningStrategy, which selects a) what Configuration to test next and b) which configuration is the best in this tuning phase. If it should not look for a new optimum it is not in a tuning phase.

Constructor & Destructor Documentation

AutoTuner()

autopas::AutoTuner::AutoTuner	(	TuningStrategiesListType &	tuningStrategies,
		const SearchSpaceType &	searchSpace,
		const AutoTunerInfo &	autoTunerInfo,
		unsigned int	rebuildFrequency,
		const std::string &	outputSuffix
	)

Constructor for the AutoTuner that generates all configurations from the given options.

Parameters

tuningStrategies	Vector of object implementing the modelling and exploration of a search space. Will be moved into the tuner.
searchSpace	All possible configurations.
autoTunerInfo	Struct containing more configuration information.
rebuildFrequency	The number of iterations after which the neighbor lists are rebuilt.
outputSuffix	Suffix for all output files produced by this object.

Member Function Documentation

addDomainSimilarityStatistics()

void autopas::AutoTuner::addDomainSimilarityStatistics	(	double	pdBinDensityStdDev,
		double	pdBinMaxDensity
	)

Adds domain similarity statistics to a vector of measurements, which can be smoothed for use in MPI Tuning to find similar domains.

Parameters

pdBinDensityStdDev	particle-dependent bin density standard deviation. See LiveInfo::gather for more information.
pdBinMaxDensity	particle-dependent bin maximum density. See LiveInfo::gather for more information.

addMeasurement()

void autopas::AutoTuner::addMeasurement	(	long	sampleRebuild,
		long	sampleTraverseParticles,
		bool	neighborListRebuilt,
		size_t	iteration,
		size_t	tuningPhase
	)

Save the runtime of a given traversal.

Samples are collected and reduced to one single value according to _selectorStrategy. Only then the value is passed on to the tuning strategy. This function expects that samples of the same configuration are taken consecutively. The sample argument is a long because std::chrono::duration::count returns a long.

Parameters

sampleRebuild	time or energy sample for rebuild part of the iteration.
sampleTraverseParticles	time or energy sample for traverse interaction part of the iteration. This includes computeInteraction and remainderTraversal call.
neighborListRebuilt	If the neighbor list as been rebuilt during the given time.
iteration	Current LogicHandler iteration.
tuningPhase	Current tuning phase number from the TuningManager

canMeasureEnergy()

bool autopas::AutoTuner::canMeasureEnergy

(

)

const

Returns whether the AutoTuner can take energy measurements.

Returns

checkEarlyStoppingCondition()

void autopas::AutoTuner::checkEarlyStoppingCondition

(

size_t

tuningPhase

)

Checks whether the current configuration performs so poorly that it shouldn't be resampled further within this tuning phase.

If the currently sampled configuration is worse than the current best configuration by more than the earlyStoppingFactor factor, it will not be sampled again this tuning phase. Uses the _estimateRuntimeFromSamples() function to estimate the runtimes.

Parameters

tuningPhase

Current tuning phase number.

forceRetune()

void autopas::AutoTuner::forceRetune

(

)

Force the internal tuner to enter a new tuning phase upon the next call to computeInteractions().

getConfigQueue()

const std::vector< Configuration > & autopas::AutoTuner::getConfigQueue

(

)

const

Getter for the current queue of configurations.

Returns

getCurrentConfig()

const Configuration & autopas::AutoTuner::getCurrentConfig

(

)

const

Get the currently selected configuration.

Returns

getEvidenceCollection()

const EvidenceCollection & autopas::AutoTuner::getEvidenceCollection

(

)

const

Getter for the internal evidence collection.

Returns

getSearchSpace()

const std::set< Configuration > & autopas::AutoTuner::getSearchSpace

(

)

const

Return the search space as a const reference.

Returns

the search space.

getTuningMetric()

const TuningMetricOption & autopas::AutoTuner::getTuningMetric

(

)

const

Getter for the primary metric for tuning.

Returns

getTuningStrategies()

const std::vector< std::unique_ptr< TuningStrategyInterface > > & autopas::AutoTuner::getTuningStrategies

(

)

const

Get the list of tuning strategies that are used.

Returns

initEnergy()

bool autopas::AutoTuner::initEnergy

(

)

Initialize pmt sensor.

Returns

True if energy measurements are enabled and possible.

inLastTuningIteration()

bool autopas::AutoTuner::inLastTuningIteration

(

)

const

Indicates if the tuner is in the last iteration of the tuning phase.

Returns

inTuningPhase()

bool autopas::AutoTuner::inTuningPhase

(

)

const

Indicate if the tuner considers itself currently in a tuning phase according to its internal counters.

Returns

logTuningResult()

void autopas::AutoTuner::logTuningResult	(	long	tuningTime,
		size_t	currentIteration
	)		const

After a tuning phase has finished, write the result to a file.

Parameters

tuningTime	Measured tuning time in nanoseconds.
currentIteration	Current LogicHandler iteration number.

needsDomainSimilarityStatistics()

bool autopas::AutoTuner::needsDomainSimilarityStatistics

(

)

const

Returns true if the AutoTuner needs domain similarity statistics.

Returns

True if the AutoTuner needs live info.

needsLiveInfo()

bool autopas::AutoTuner::needsLiveInfo

(

)

const

Returns true if the AutoTuner needs live info.

Returns

True if the AutoTuner needs live info.

operator=()

AutoTuner & autopas::AutoTuner::operator= ( AutoTuner &&  other )

noexcept

Move assignment operator.

Parameters

other

Returns

receiveLiveInfo()

void autopas::AutoTuner::receiveLiveInfo	(	const LiveInfo &	liveInfo,
		bool	isStartOfTuningPhase
	)

Pass live info on to all tuning strategies.

Parameters

liveInfo
isStartOfTuningPhase	True, if the next iteration starts a new tuning phase.

rejectConfig()

Configuration autopas::AutoTuner::rejectConfig	(	const Configuration &	rejectedConfig,
		bool	indefinitely,
		size_t	tuningPhase
	)

Tell the tuner that the given config is not applicable.

Since this operation might change the suggestion what configuration to try next, this next suggestion is returned.

Note

The applicability checking logic was moved out of the tuner because it needed the container, thus raising the compile complexity.

Parameters

rejectedConfig
indefinitely	Whether the given config should be completely removed from the search space (aka rejected indefinitely).
tuningPhase	Current tuning phase.

Returns

Tuple<Next configuration to use, still tuning>.

resetEnergy()

bool autopas::AutoTuner::resetEnergy

(

)

Reset the rapl meter to prepare for a new measurement.

Returns

True if energy measurements are possible on this system.

sampleEnergy()

std::tuple< double, double, double, long > autopas::AutoTuner::sampleEnergy

(

)

Take an energy measurement.

Returns

Tuple<PsysEnergy, PkgEnergy, RamEnergy, TotalEnergy>

searchSpaceIsEmpty()

bool autopas::AutoTuner::searchSpaceIsEmpty

(

)

const

Indicator function whether the search space has no configurations in it.

Returns

searchSpaceIsTrivial()

bool autopas::AutoTuner::searchSpaceIsTrivial

(

)

const

Indicator function whether the search space consists of exactly one configuration.

Returns

setOptimalConfiguration()

void autopas::AutoTuner::setOptimalConfiguration

(

const Configuration &

optimalConfig

)

Pushes the provided config into the _configQueue and ends tuning immediately.

Used by the TuningManager, when we want a different configuration than the one determined by one AutoTuner in isolation.

Parameters

optimalConfig

The configuration to be used from now on until the next tuning phase.

setRebuildFrequency()

void autopas::AutoTuner::setRebuildFrequency

(

double

rebuildFrequency

)

Sets the _rebuildFrequency.

This is the average number of iterations per rebuild. This is used to dynamically change the _rebuildFrequency based on estimate in case of dynamic containers.

Parameters

rebuildFrequency

Current rebuild frequency in this instance of autopas, used by autopas for weighing rebuild and non-rebuild iterations

tuneConfiguration()

bool autopas::AutoTuner::tuneConfiguration	(	size_t	currentIteration,
		size_t	tuningPhase,
		bool	isStartOfTuningPhase
	)

Tune available algorithm configurations.

It is assumed this function is only called for relevant functors and that at least two configurations are allowed. When in tuning phase selects next config to test. At the end of the tuning phase select optimum. The function returns true if the selected config is not yet the optimum but something that should be sampled.

Parameters

currentIteration	Current LogicHandler iteration number.
tuningPhase	Current tuning phase number.
isStartOfTuningPhase	True if this is the start of a new tuning phase.

Returns

true iff still in tuning phase.

willRebuildNeighborLists()

bool autopas::AutoTuner::willRebuildNeighborLists

(

)

const

Returns whether rebuildNeighborLists() should be triggered in the next iteration.

This indicates a configuration change. In the non-tuning phase, the rebuildNeighborLists() is triggered in LogicHandler.

Returns

True if the current iteration counters indicate a rebuild in the next iteration due to a configuration change.

---

The documentation for this class was generated from the following files:

• autopas/tuning/AutoTuner.h
• autopas/tuning/AutoTuner.cpp